Machine learning in trading: theory, models, practice and algo-trading - page 1804

 
Igor Makanu:

I'm asking you - did you try the other side? She doesn't know to the left, does she? - You and I know that the graph moves to the right, don't tell her!)

Tried it, no difference. But if the test is to the right, then it is too late when it starts to trade. It's better to make a trace on recent data, imha, then the probability of profit is higher.


 
Maxim Dmitrievsky:

It is better to trace on the latest data, imha, then the probability of profit is higher

yes it is clear

If you can determine the lifetime of the TS - if you can, then everything is ok - trade through its time - teach it - trade through - teach it...

When I look at it that way there is too little data for the test, that's why I suggest checking both right and left side, while it makes no sense to look deep into the history, volatility changes all the time, that's why only a couple of years are relevant, at least I use genetics for 1.5 year options, forwardtest for 6 months, sometimes for a year, more than a year the TS usually have big drawdowns - not the one I would like

 
Igor Makanu:

It's clear.

If you can determine the durability time of the TS - if you can, then everything is OK - traded its time - taught - traded - taught ...

I don't have enough data for the test, that's why I suggest checking both the left and the right side, while it makes no sense to look deep into the history, volatility changes all the time, that's why only a couple of years are relevant, at least I use genetics for 1.5 year options, I search forwardtest for 6 months, sometimes for a year, more than a year the TS appears with big drawdowns - not the one I would like

I have no qualitative signs, I have to make up a bicycle. If the 5X will reject the trane, it is anyway a good model.

 
Maxim Dmitrievsky:

There are no quality signs, you have to make up a bicycle. If the 5X from the trane is a good model at any rate

well, that's the problem, do not guess the future - how long it will take, and plan

if the total volume of lots has also increased, then they will probably end up draining at comet speed ))))) At least when I look at the signals, I don't know if it's a good model.


If I look at the signals, I saw a smooth chart with small drawdowns, and then the arrow goes down, starting to watch the transactions, everything is clear, the TS has stopped working, and the drawdown is much higher than when starting the TS.... as if the anti-Martingale triumphs - that was an old topic ;)

 
Igor Makanu:

Well, that's the problem, do not guess the future - how long it will last, and plan it

If you do not know the future - how long will it take to deposit the funds in the chart, then let's see if the loss is beginning, and if the lot has also increased, then let's go down the same way as the comet does)))


If I look at the signals, I saw a smooth chart with small drawdowns, and then the arrow goes down, when you start watching the transactions, everything is clear, the TS stopped working, and the yield is much higher than at the start of TS.... kind of like the anti-marthingale triumphs - there was an old topic ;)

Anyway, if you have any ideas what can be screwed in as chips, I can very quickly check )

I've tried both increments and regressions, the results are from case to case

They never made the news in the tester, it would have been fun to play with them
 
Maxim Dmitrievsky:


2 years OOS, training last 5 months.

To put a period of 2 years). A lot depends on the periods it turns out. And models to impose, without them apparently nowhere. The question of the minimum period of reliability.

 
Valeriy Yastremskiy:

(I'm not sure about the period of two years). A lot depends on the periods. And modeling to superimpose, without them apparently nowhere. The question is the minimum period of reliability.

On Eurobucks almost all features are trained well for 2-3 years, there is clearly a pattern. It is worse with other instruments.

The bot really finds the pattern, if there is one. Features do not play a special role.

 
Maxim Dmitrievsky:


the news in the tester was not made, it would be fun to play with them

News if only to parse the downloaded file, there are for 10 years likehttps://www.forexfactory.com/showthread.php?t=549773

But yes, I have long wanted to quickly and easily scour the news)

 
Valeriy Yastremskiy:

News if only to parse the downloaded file, there are for 10 years likehttps://www.forexfactory.com/showthread.php?t=549773

But yes, for a long time I wanted to quickly and easily scour the news)

do not want to bother, synchronize with the data tester ... Bae :) although ... what the hell

news in the tester mast hav
 
Maxim Dmitrievsky:

On eurobucks almost any chip for 2-3 years is well trained, there is clearly a pattern. With other instruments it is worse.

It is also necessary to be trained on the instrument). If the matrix changes during a period shorter than the period necessary for learning then it seems to be SB.)